After-hours customer enquiries: designing an AI assistant that knows when to hand off

Start with one job your team can describe

Begin with a narrow outcome: answer common service questions and prepare callback requests. List the information a staff member actually uses for that job. A service catalogue, coverage areas, working hours, and approved pricing explanations are more useful than an unfiltered folder of old documents.

Give each source an owner and a review date. Mark temporary offers and exceptions clearly. When two documents disagree, decide which one takes precedence. An assistant should not invent a missing rule simply because the conversation would otherwise pause.

Separate an answer from a business commitment

Use a simple decision table before writing conversation scripts. These are illustrative rules for a home-services company, not claims about an existing Arczen client. Adapt them to your actual operations and give staff the same rules.

Make the handoff an operational step

A message saying “I will pass this on” is not evidence that anyone received the request. Google Cloud’s Dialogflow CX documentation illustrates the distinction: its live-agent handoff response provides a signal, while the receiving system or integration must perform the handoff actions.

For your business, identify the destination: a shared inbox, a ticket queue, or a customer record with a named owner. Define who checks it when work resumes and who covers absences. If no one is available now, offer a callback request with your actual response window. Do not label that experience a live transfer.

• Include the customer’s request, preferred contact method, relevant facts, and unresolved question.
• Record what the assistant already said, especially any quoted fee or promised follow-up.
• Confirm successful receipt before telling the customer their request was recorded; show an alternative contact route if delivery fails.

References: Google Cloud: Dialogflow CX fulfillments and live-agent handoff

An illustrative after-hours conversation

Imagine a customer asking a home-maintenance business: “Can you repair my air conditioner tomorrow morning?” The assistant can ask for the service area and preferred contact method, then explain that the team must confirm the appointment. It should not turn a preferred time into a booked slot.

Once the request is stored, an appropriate reply could be: “Your callback request has been received. You asked for tomorrow morning; that appointment is not confirmed. Our team will contact you during the response window shown above.” The wording must match the business’s actual staffing and the system’s confirmed result.

If the customer asks to speak to someone immediately, acknowledge that request directly. Avoid making them repeat the same details through several menus before revealing that the team is offline.

Rehearse the conversations that can go wrong

Create a small acceptance checklist with the people who will handle requests. Write the expected outcome before running each conversation. Where you serve Arabic and English speakers, test both languages and customers switching between them.

1. Ask a routine question with a documented answer. Check that the answer preserves conditions and exclusions.
2. Ask an ambiguous question or request an unavailable service. Expect a useful clarification or handoff, not a guess.
3. Change contact details midway through the conversation. Verify that the final request contains the corrected information.
4. In a test environment, make the booking or ticket connection fail. Check the customer message and the team’s failure alert.
5. Request a person, use frustrated language, and ask again after a misunderstanding. Check that escalation remains easy.

Review what reaches the team

Start with a limited service scope and review a sample of conversations regularly. Track unanswered questions, incorrect answers, missing contact details, failed deliveries, and requests waiting for an owner. A fast first reply alone does not show whether the customer received useful help.

Collect only the information needed for the agreed task. Decide who can see conversation records and how long you need them. Use what staff learn from real follow-ups to improve the approved answers and handoff rules before expanding the assistant’s responsibilities.